See Different, Think Better: Visual Variations Mitigating Hallucinations in LVLMs

ViHallu uses visual variations and visual instruction construction for visual-semantic alignment.

Bridging Perception and Reasoning: Token Reweighting for RLVR in Multimodal LLMs

ToR reweights coupled perception and reasoning tokens during multimodal RLVR.

Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models

PGPO amplifies learning signals for visually dependent tokens.

Entropy Is Not Enough: Unlocking Effective Reinforcement Learning for Visual Reasoning via Vision-Anchored Token Selection

VEPO combines visual sensitivity and token entropy for visual reasoning RL.

Adversarial Orthogonal Disentanglement for LVLM Hallucination Mitigation

AOD disentangles hallucination directions for training-free contrastive decoding in LVLMs.

Search-G1: Grounded Search Agents via Representation-Based Intrinsic Rewards

Grounded search agents; public arXiv record is pending.

Hallucination-aware intermediate representation edit in large vision-language models

Intermediate representation editing for LVLM hallucination mitigation.

NoLan: Mitigating Object Hallucinations in Large Vision-Language Models via Dynamic Suppression of Language Priors

Dynamic suppression of language priors during decoding.

Understanding and Mitigating Hallucinations in Multimodal Chain-of-Thought Models

Hallucination analysis and mitigation for multimodal CoT.

R-CoV: Region-Aware Chain-of-Verification for Alleviating Object Hallucinations in LVLMs

Region-aware verification loops for object hallucinations.

Thinking in Uncertainty: Mitigating Hallucinations in MLRMs with Latent Entropy-Aware Decoding

Entropy-aware decoding for multimodal reasoning models.

Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning

Constructs a robust instruction tuning dataset with positive and negative samples to update the model.

Aligning Large Multimodal Models with Factually Augmented RLHF

Introduces a factually augmented mechanism, using reinforcement learning from human feedback to align the model.

Volcano: Mitigating Multimodal Hallucination through Self-Feedback Guided Revision

Specifically trains the MLLM to generate self-feedback and revise its output based on it.

HalluciDoctor: Mitigating Hallucinatory Toxicity in Visual Instruction Data

Cleans "hallucinatory toxicity" in existing visual instruction datasets and generates counterfactual data for fine-tuning.

Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization

Constructs hallucination-aware data pairs and applies the DPO algorithm directly to update model weights.

RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback

Collects fine-grained, segment-level human correctional feedback for reward modeling and PPO optimization.

ViGoR: Improving Visual Grounding of Large Vision Language Models with Fine-Grained Reward Modeling

Trains a fine-grained visual reward model to guide the model toward better visual alignment.

Visually Dehallucinative Instruction Generation: Know What You Don't Know

Teaches the model to actively output "I don't know" when evidence is insufficient via specialized instruction tuning.

Aligning Modalities in Vision Large Language Models via Preference Fine-tuning

Constructs a preference-based dataset to align the model with high-quality image descriptions via preference fine-tuning.

Less is More: Mitigating Multimodal Hallucination from an EOS Decision Perspective

Reconstructs the fine-tuning dataset to teach the model to output the End-Of-Sequence (EOS) token early to truncate generation.

Mitigating Dialogue Hallucination for Large Vision Language Models via Adversarial Instruction Tuning

Introduces adversarial instruction tuning to improve model robustness against misleading user queries.

Calibrated Self-Rewarding Vision Language Models

Proposes a self-calibrating reward mechanism enabling iterative self-generated feedback and preference alignment.

Reflective Instruction Tuning: Mitigating Hallucinations in Large Vision-Language Models

Constructs a reflective instruction dataset to teach the model to perform internal visual fact-checking before outputting answers.

Investigating and Mitigating Object Hallucinations in Pretrained Vision-Language (CLIP) Models

Cleans and rewrites captions in CLIP pre-training data to reduce object hallucinations at the vision-language source.

V-DPO: Mitigating Hallucination in Large Vision Language Models via Vision-Guided Direct Preference Optimization

Uses a vision-guided mechanism to construct high-quality positive and negative preference pairs for DPO.

Hallucination-resistant multimodal content generation through knowledge graph-based reinforcement learning

Grounds multimodal generation in structured knowledge and optimizes graph-based rewards to suppress unsupported content.

Adaptive Hallucination Alleviation in Multimodal Large Language Models: From Strategic Data Selection to Severity-Guided Training

Selects hallucination-oriented samples, assigns severity-specific loss weights, and modulates localized visual attention during HD-DPO.

EchoBat: Echo-Vision Enhancement and Echo-Layered Sampling for Video LLMs Hallucination Mitigation

Builds temporally grounded video preferences from visual and audio evidence and improves supervision with echo-layered keyframe sampling.

VGL-DPO: Vision-Guided Lexical Direct Preference Optimization for Mitigating Hallucination in Multimodal Large Language Models

Reweights positive words by visual relevance and adapts the negative preference loss using lexical importance differences.

Multimodal Chain-of-Thought Reasoning in Language Models

Adopts a two-stage fine-tuning framework, training the model to generate a CoT reasoning process before the answer.

Hallucination Augmented Contrastive Learning for Multimodal Large Language Model

Introduces a hallucination-augmented contrastive learning loss function during the training phase.

KAM-CoT: Knowledge Augmented Multimodal Chain-of-Thoughts Reasoning

Jointly trains a language model, a vision encoder, and a Graph Neural Network (GNN) to integrate knowledge graphs for reasoning.

Plausible May Not Be Faithful: Probing Object Hallucination in Vision-Language Pre-training

Proposes a novel vision-language pre-training (VLP) objective to suppress hallucinations.

Mitigating Fine-Grained Hallucination by Fine-Tuning Large Vision-Language Models with Caption Rewrites

Uses ChatGPT to rewrite image captions for data construction and performs two-stage fine-tuning.

Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback

Proposes hallucination-aware DPO, utilizing fine-grained AI feedback to construct positive and negative pairs for weight updates.

RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness

Utilizes fine-grained feedback from open-source AIs to build a preference dataset and updates weights using alignment algorithms.

mDPO: Conditional Preference Optimization for Multimodal Large Language Models

Proposes a conditional preference optimization algorithm to reduce hallucinations without degrading general capabilities.

CLIP-DPO: Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs

Uses a pre-trained CLIP model directly as an AI judge to generate preference pairs for DPO training.

EFUF: Efficient Fine-grained Unlearning Framework for Mitigating Hallucinations in Multimodal Large Language Models

Proposes an efficient fine-grained unlearning framework to "forget" the tendency to hallucinate via reverse parameter updates.

Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization

Introduces a hallucination-inducing mechanism during training to construct sample pairs for parameter optimization.

Mitigating Open-Vocabulary Caption Hallucinations

Trains a reward model based on Natural Language Inference (NLI) and uses RL to fine-tune the MLLM.

Prescribing the Right Remedy: Mitigating Hallucinations in Large Vision-Language Models via Targeted Instruction Tuning

Constructs targeted repair instruction datasets for different hallucination types to perform targeted fine-tuning.

See or Guess: Counterfactually Regularized Image Captioning

Introduces a counterfactual regularization term during training to force distinction between seen visuals and guessed priors.

Mitigating Hallucination in Multimodal Large Language Model via Hallucination-targeted Direct Preference Optimization

Generates targeted preference data specifically for hallucinations and applies DPO to penalize hallucination-prone features.

Silkie: Preference Distillation for Large Visual Language Models

Uses AI to construct the VLFeedback dataset and distills preferences into the model via DPO.

Mitigating Object Hallucination in MLLMs via Data-augmented Phrase-level Alignment

Constructs hard negative samples via data augmentation and introduces a contrastive loss to strengthen visual grounding.

Mitigating Hallucination in Visual Language Models with Visual Supervision

Constructs a fine-grained RAI-30k dataset and integrates the SAM model during instruction tuning.

Mitigating Multilingual Hallucination in Large Vision-Language Models

Builds a correctional dataset specifically for multilingual scenarios to mitigate cross-lingual hallucinations caused by language bias.

Modality-Fair Preference Optimization for Trustworthy MLLM Alignment

Proposes a modality-fair preference optimization algorithm to balance penalties and prevent language prior dominance.

Token Preference Optimization with Self-Calibrated Visual-Anchored Rewards for Hallucination Mitigation

Proposes a token-level preference optimization strategy based on self-calibrated visual-anchored rewards.

Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization

Introduces Retrieval-Augmented Generation (RAG) into preference data construction to guide DPO training.

Symmetrical Visual Contrastive Optimization: Aligning Vision-Language Models with Minimal Contrastive Images

Proposes symmetrical visual contrastive optimization using minimal contrastive images for alignment fine-tuning.

FGAIF: Aligning Large Vision-Language Models with Fine-grained AI Feedback

Uses fine-grained AI feedback to construct preference data and applies DPO for alignment training.

TextSquare: Scaling up Text-Centric Visual Instruction Tuning

Massively scales up text-centric visual instruction tuning data to specifically address text hallucinations in OCR tasks.

Understanding Alignment in Multimodal LLMs: A Comprehensive Study

Systematically builds datasets covering various alignment strategies to empirically verify their impact on hallucinations.

RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data

Optimizes a visual reward model using auxiliary text-only preference data during RLHF to enhance hallucination discrimination.

HELPD: Mitigating Hallucination of LVLMs by Hierarchical Feedback Learning with Vision-enhanced Penalty Decoding

Proposes a fine-tuning framework combining vision-enhanced penalty decoding with hierarchical feedback learning for behavior alignment.

Mitigating Hallucinations in Large Vision-Language Models via Entity-Centric Multimodal Preference Optimization

Proposes entity-centric multimodal preference optimization, focusing penalty granularity precisely on specific visual entity tokens.

Fact :Teaching MLLMs with Faithful, Concise and Transferable Rationales

Introduces faithful and concise reasoning rationales as supervision signals during training to enhance logical generation.

Generating Faithful and Salient Text from Multimodal Data

Constructs a multimodal factual consistency dataset and fine-tunes the model to improve faithfulness in data-to-text generation.

Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment

Formulates preference modeling as a next-token prediction task to update the model using fine-grained verifier data.

Verb Mirage: Unveiling and Assessing Verb Concept Hallucinations in Multimodal Large Language Models

Constructs a dataset targeting verb concept hallucinations and applies specific instruction tuning to correct action biases.

Beyond Human Data: Aligning Multimodal Large Language Models by Iterative Self-Evolution

Automatically generates and filters high-quality alignment data through an iterative self-evolution mechanism.

CHiP: Cross-modal Hierarchical Direct Preference Optimization for Multimodal LLMs

Proposes cross-modal hierarchical DPO to penalize global image-text matching and local object features at different levels.

PerturboLLaVA: Reducing Multimodal Hallucinations with Perturbative Visual Training

Introduces subtle visual perturbations during training forward passes to force the learning of robust visual representations.

Mitigating Low-Level Visual Hallucinations Requires Self-Awareness: Database, Model and Training Strategy

Builds a low-level visual hallucination database and applies negative sampling to enhance awareness of low-level features.

Critique Before Thinking: Mitigating Hallucination through Rationale-Augmented Instruction Tuning

Adopts rationale-augmented instruction tuning to teach the model to generate critique logic before providing an answer.

OViP: Online Vision-Language Preference Learning for VLM Hallucination

Proposes an online preference learning framework to dynamically generate preference pairs during training.

BIMA: Bijective Maximum Likelihood Learning Approach to Hallucination Prediction and Mitigation in Large Vision-Language Models

Employs a bijective maximum likelihood learning approach to suppress hallucinations by optimizing joint probability distributions.

Stop learning it all to mitigate visual hallucination, Focus on the hallucination target

Stops learning redundant backgrounds and forces weight updates to focus exclusively on target areas causing hallucinations.

Identify, Isolate, and Purge: Mitigating Hallucinations in LVLMs via Self-Evolving Distillation

Identifies and purges hallucinated samples from fine-tuning data, using the refined high-quality data for knowledge distillation.

ReLoop: "Seeing Twice and Thinking Backwards" via Closed-loop Training to Mitigate Hallucinations in Multimodal understanding

Proposes a closed-loop framework where the model generates predictions, performs backward verification, and uses errors as training signals.

Analyzing and Mitigating Object Hallucination: A Training Bias Perspective

Analyzes object biases introduced during pre-training and debiases the model using a reverse objective function during fine-tuning.

TARS: MinMax Token-Adaptive Preference Strategy for Hallucination Reduction in MLLMs

Adopts an adaptive MinMax Token preference strategy to dynamically adjust reward allocation.

Mitigating Hallucinations in Multimodal LLMs via Object-aware Preference Optimization

Integrates the CHAIR metric, utilizing object-level matching scores as the preference reward for DPO.

Two Causes, Not One: Rethinking Omission and Fabrication Hallucinations in MLLMs

Distinguishes "omission" and "fabrication" hallucination causes and constructs respective fine-tuning data for decoupled training.

Mitigating Visual Hallucinations via Semantic Curriculum Preference Optimization in MLLMs

Proposes semantic curriculum preference optimization, training the model in increasing difficulty to smoothly eliminate hallucinations.

Intervene-All-Paths: Unified Mitigation of LVLM Hallucinations across Alignment Formats

Proposes a universal training framework across multiple alignment formats (e.g., DPO, PPO) to eliminate vision-language biases.

OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation

Extracts self-attention matrices to find and penalize attention flows overly reliant on summary tokens.

Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attention

Fuses global and local attention feature matrices internally to enhance fine-grained perception.

Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs

Adjusts cross-modal attention by forcefully amplifying the weights of visual tokens.

DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination

Dives into the MLLM to mask specific attention connections overly dominated by text.

Mitigating Modality Prior-Induced Hallucinations in Multimodal Large Language Models via Deciphering Attention Causality

Uses causal analysis to block specific causal attention paths causing hallucinations during inference.

Mitigating Object Hallucination via Concentric Causal Attention

Strengthens the causal association between local visual features and global semantics in attention matrices.

Nullu: Mitigating Object Hallucinations in Large Vision-Language Models via HalluSpace Projection

Projects hidden states into a hallucination space and strips them via orthogonalization to clear hallucinations.

VASparse: Towards Efficient Visual Hallucination Mitigation via Visual-Aware Token Sparsification

Introduces visual-aware sparsification to directly drop irrelevant visual tokens prone to inducing hallucinations.

Mitigating Hallucinations in Large Vision-Language Models by Adaptively Constraining Information Flow

Adaptively constrains inter-layer information flow to block false causal paths causing object hallucinations.

ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large language Models

Amplifies low-level visual signals in multimodal layers to combat dilution by deep language features.

MCA-LLaVA: Manhattan Causal Attention for Reducing Hallucination in Large Vision-Language Models

Reconstructs multimodal attention using a causal matrix based on Manhattan distance.

Diving into Mitigating Hallucinations from a Vision Perspective for Large Vision-Language Models

Intervenes from the visual processing architecture to fundamentally alter how the model acquires visual information.

Scalpel: Fine-Grained Alignment of Attention Activation Manifolds via Mixture Gaussian Bridges to Mitigate Multimodal Hallucination

Maps hallucination activations toward trusted attention regions with mixture-Gaussian bridges.

RFI: Rectified Flow Intervention for Mitigating Object Hallucination in Large Vision-Language Models

Predicts input-specific latent steering vectors with gradient correction in a single forward pass.

Taming the Phantom: Token-Asymmetric Filtering for Hallucination Mitigation in Large Vision-Language Models

Filters intermediate attention maps to isolate dominant phantom text tokens and strengthen visual anchor tokens.

Look Before You Speak: Visual Re-Focusing for Hallucination Mitigation in Large Vision-Language Models

Reorients the model toward image evidence before response generation to reduce visually unsupported claims.

Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens

Uses an attention lens to locate hallucinated features in middle layers and masks them for correction.

Cracking the Code of Hallucination in LVLMs with Vision-aware Head Divergence

Quantifies perception divergence and dynamically masks specific heads that lose visual alignment.

Attention Reallocation: Towards Zero-cost and Controllable Hallucination Mitigation of MLLMs

Reallocates Attention Maps during inference to pull misplaced attention back to visual targets.

ICT: Image-Object Cross-Level Trusted Intervention for Mitigating Object Hallucination in Large Vision-Language Models

Proposes cross-level trusted intervention to block the drift of object representations toward false language priors.

Who Brings the Frisbee: Probing Hidden Hallucination Factors in Large Vision-Language Model via Causality Analysis

Implements causal intervention to cut off spurious feature maps hidden in attention matrices.

From Pixels to Tokens: Revisiting Object Hallucinations in Large Vision-Language Models

Locates the evolution of object features and corrects visual tokens via early layer interventions.

Seeing Clearly by Layer Two: Enhancing Attention Heads to Alleviate Hallucination in LVLMs

Accurately locates key attention heads and specifically enhances or suppresses them during forward propagation.

Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models

Calibrates the weight allocation of attention heads to weaken the implicit coverage of visual inputs by language priors.

Looking Beyond Text: Reducing Language bias in Large Vision-Language Models via Multimodal Dual-Attention and Soft-Image Guidance

Introduces dual-attention mechanisms to internally rebalance language priors and image feature flows.

CLAIM: Mitigating Multilingual Object Hallucination in Large Vision-Language Models with Cross-Lingual Attention Intervention

Cuts off implicit semantic biases introduced during cross-lingual transitions inside the model.

HalLoc: Token-level Localization of Hallucinations for Vision Language Models

Pinpoints tokens sourcing hallucinations and excises them during forward propagation.

Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations

Identifies and edits vision-language feature vectors representing hallucinations in intermediate layers.

Reducing Hallucinations in Vision-Language Models via Latent Space Steering

Injects pre-trained steering vectors into the latent space to guide generation away from hallucinations.

Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding

Penalizes decoding by contrasting the logit output distributions of original and distorted images during inference.

Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance

Introduces Classifier-Free Guidance (CFG) during inference to amplify the influence of visual features in the generation process.

IBD: Alleviating Hallucinations in Large Vision-Language Models via Image-Biased Decoding

Penalizes pure language priors by contrasting the logits of inputs with and without images.

HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding

Proposes adaptive focal-contrast decoding to fine-grainedly suppress hallucinations at the local token level.

Multi-Modal Hallucination Control by Visual Information Grounding

Introduces a visual grounding mechanism during decoding to correct biases by contrasting different information streams.

Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding

Constructs positive and negative instructions, contrasting their logits during inference to guide decoding.

Investigating and Mitigating the Multimodal Hallucination Snowballing in Large Vision-Language Models

Introduces residual visual connections to add early visual features directly into subsequent decoding computations.

ConVis: Contrastive Decoding with Hallucination Visualization for Mitigating Hallucinations in Multimodal Large Language Models

Uses attention visualization to locate hallucination regions and constructs interference images for contrastive decoding.

Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models

Introduces a generative feedback loop to detect and correct the probability distribution of the current token in real time.

Octopus: Alleviating Hallucination via Dynamic Contrastive Decoding

Adaptively adjusts the contrastive penalty based on image complexity and local token features.

Mixture of Decoding: An Attention-Inspired Adaptive Decoding Strategy to Mitigate Hallucinations in Large Vision-Language Models

Adaptively switches between different sampling algorithms based on attention strength.

Med-VCD: Mitigating hallucination for medical large vision language models through visual contrastive decoding

Selects visually informed tokens on the fly, removing redundant tokens while retaining critical medical image context.

Mitigating Hallucinations in Large Vision-Language Models (LVLMs) via Language-Contrastive Decoding (LCD)

Constructs an image-free text input as a negative reference and subtracts language prior scores during logit calculation.

CODE: Contrasting Self-generated Description to Combat Hallucination in Large Multi-modal Models

Contrasts logit distributions based on the original image versus self-generated descriptions to penalize text-reliant tokens.

EventHallusion: Diagnosing Event Hallucinations in Video LLMs

Suppresses event hallucinations by contrasting output probabilities between full and truncated videos.

Seeing is Believing: Mitigating Hallucination in Large Vision-Language Models via CLIP-Guided Decoding

Introduces an external CLIP model to score and guide candidate tokens during decoding to improve image-text consistency.

Mitigating Hallucination in Visual-Language Models via Re-Balancing Contrastive Decoding

Dynamically adjusts the logit weight distribution of text and image features during inference.

Mitigating Hallucinations in Large Vision-Language Models with Internal Fact-based Contrastive Decoding

Extracts internally generated facts as references for contrastive decoding to suppress over-extrapolation.

HallE-Control: Controlling Object Hallucination in Large Multimodal Models

Introduces a control parameter during decoding to force switching between pure contextual description and parameterized knowledge imagination.

Sycophancy in Vision-Language Models: A Systematic Analysis and an Inference-Time Mitigation Framework

Introduces language contrastive decoding to weaken misleading priors from user prompts to combat sycophancy-induced hallucinations.

Mitigating Hallucinations in Large Vision-Language Models via Summary-Guided Decoding

Utilizes image summaries as guidance signals to intervene in output probabilities for contextual consistency.

Retrieval Visual Contrastive Decoding to Mitigate Object Hallucinations in Large Vision-Language Models

Contrasts logits from the original image and retrieved similar reference images to highlight core facts.

CoFi-Dec: Hallucination-Resistant Decoding via Coarse-to-Fine Generative Feedback in Large Vision-Language Models

Generates multi-granularity hierarchical feedback to dynamically correct outputs.

Analyzing and Mitigating Object Hallucination in Large Vision-Language Models

Trains an independent Reviser model specifically to detect and reconstruct hallucinated entities in outputs.

Woodpecker: Hallucination Correction for Multimodal Large Language Models

Calls external visual foundation models to verify entity facts, finally using an LLM to rewrite and correct.

Visual Fact Checker: Enabling High-Fidelity Detailed Caption Generation

Constructs a pipeline system containing claim extraction, external detection verification, and corrected generation.

Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning

Modifies reference images to expose hallucinating agents and replaces consensus-only debate with evidence-based factual verification.

Mitigating Object and Relationship Hallucination in Large Vision Language Model with Multi-Agent Guidance

Coordinates specialized agents to identify and correct unsupported object and relationship claims.

CEBC: Conformal Evidence-Bounded Control for Low-Hallucination Vision-Language Generation

Calibrates an external detector on held-out data and minimally revises unsupported object mentions under explicit risk bounds.

Logical Closed Loop: Uncovering Object Hallucinations in Large Vision-Language Models

Uses external language models to ask questions about generated claims for self-verification.

Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning

Utilizes a bottom-up holistic reasoning framework combined with multi-perspective prompt verification.

Reefknot: A Comprehensive Benchmark for Relation Hallucination Evaluation, Analysis and Mitigation in Multimodal Large Language Models

Uses an external detector to locate relation errors, then applies template-based secondary correction.

Pensieve: Retrospect-then-Compare Mitigates Visual Hallucination

Adopts a "Retrospect-then-Compare" multi-step prompt to make the model review visual details for correction.

Mitigating Hallucinations in Multimodal Spatial Relations through Constraint-Aware Prompting

Explicitly adds spatial relation logic rules into the prompt to correct spatial confusion.

RITUAL: Random Image Transformations as a Universal Anti-hallucination Lever in Large Vision Language Models

Applies random geometric transformations to construct multiple views for consistency verification and correction.

Systematic Reward Gap Optimization for Mitigating VLM Hallucinations

The model generates a draft, then independently reviews and self-scores sub-topics for correction.

InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration

Deploys a network of introspective and external verification agents for interactive debate.

Alleviating Hallucination in Large Vision-Language Models with Active Retrieval Augmentation

Evaluates model uncertainty to actively trigger knowledge base retrieval for entity correction.

A Unified Hallucination Mitigation Framework for Large Vision-Language Models

Builds a unified cross-modal diagnosis pipeline utilizing external discriminators to intercept and modify outputs.

Mitigating Object Hallucinations via Sentence-Level Early Intervention

Implements factual truncation prompts in early sentence stages of generation to prevent error propagation.

Exploring the Transferability of Visual Prompting for Multimodal Large Language Models

Overlays visual prompts (e.g., circles, bounding boxes) on the original image to force model attention.

What if...?: Thinking Counterfactual Keywords Helps to Mitigate Hallucination in Large Multi-modal Models

Introduces "What if" prompting to compel the model to self-reflect on its initial judgments.

From Training-Free to Adaptive: Empirical Insights into MLLMs’ Understanding of Detection Information

Feeds information identified by external object detection models directly to the MLLM as context.

Visual Description Grounding Reduces Hallucinations and Boosts Reasoning in LVLMs

Pre-supplements missing low-level visual perception details in the input prompt to bridge the gap.

Do More Details Always Introduce More Hallucinations in LVLM-based Image Captioning?

Strictly controls the requested detail quantity via prompt to balance richness and hallucination rate.

Look, Compare, Decide: Alleviating Hallucination in Large Vision-Language Models via Multi-View Multi-Path Reasoning

Adopts multi-view prompts and multi-path reasoning, aggregating and comparing answers.

Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models

Uses prompts to guide the model to actively re-extract visual clues for confirmation during generation.

Scaling Inference-Time Search with Vision Value Model for Improved Visual Comprehension

Introduces an external Vision Value Model to assist tree search, guiding autoregressive generation.

Black-Box Visual Prompt Engineering for Mitigating Object Hallucination in Large Vision Language Models

Uses black-box optimization to automatically search for "visual prompt patches" that suppress hallucinations.

Entropy-optimized contrastive decoding for hallucination suppression in vision-language-action models

Uses uncertainty-aware contrastive calibration to suppress hallucinated actions in vision-language-action generation.

VCGD: Visual Clue Guided Decoding with Caption Model for Mitigating Hallucination in Multimodal Large Language Models

Uses a reinforcement-learned caption model as an auxiliary visual clue source and applies image-confidence constraints during generation.

CTDD: Cumulative Trend Divergence Decoding for Mitigating Hallucination in Large Vision-Language Models

Tracks accumulated divergence in token-distribution trends and calibrates generation when visual and linguistic evidence separate.

Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models

Filters hallucinated words by comparing the confidence of different generation paths via the model's own introspective mechanism.

HaloQuest: A Visual Hallucination Dataset for Advancing Multimodal Reasoning

Evaluates synthetic images / vqa with Gen tasks using Acc metrics.

HalluAudio: A Comprehensive Benchmark for Hallucination Detection in Large Audio-Language Models

Evaluates speech/sound/music with Dis & Gen tasks using Acc / Hallucination Rate metrics.

Detecting and Evaluating Medical Hallucinations in Large Vision Language Models

Evaluates medical context with Dis & Gen tasks using MediHall Score metrics.

VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models

Evaluates temporal/extrinsic with Dis & Gen tasks using Acc metrics.

MedVH: Towards Systematic Evaluation of Hallucination for Large Vision Language Models in the Medical Context

Evaluates medical context with Dis & Gen tasks using Characterization Score metrics.

Localizing Before Answering: A Hallucination Evaluation Benchmark for Grounded Medical Multimodal LLMs

Evaluates grounded medical with Gen tasks using Acc / Loc metrics.

MIHBench: Benchmarking and Mitigating Multi-Image Hallucinations in Multimodal Large Language Models

Evaluates multi-image reasoning with Dis & Gen tasks using Acc metrics.

The Instinctive Bias: Spurious Images lead to Illusion in MLLMs

Evaluates spurious images with Dis tasks using Acc metrics.

Explore the Hallucination on Low-level Perception for MLLMs

Evaluates low-level perception with Gen tasks using Acc metrics.

JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images

Evaluates imaginary vqa with Dis & Gen tasks using Acc metrics.

Detecting and Preventing Hallucinations in Large Vision Language Models

Evaluates fine-grained with Dis tasks using Reward Score metrics.

Negative Object Presence Evaluation (NOPE) to Measure Object Hallucination in Vision-Language Models

Evaluates negative pronoun with Dis tasks using Acc/METEOR metrics.

THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models

Evaluates free-form generations with Gen tasks using P/R/F metrics.

Quantity Matters: Towards Assessing and Mitigating Number Hallucination in Large Vision-Language Models

Evaluates object counting with Dis tasks using Acc / Consistency metrics.

CIEM: Contrastive Instruction Evaluation Method for Better Instruction Tuning

Evaluates multimodal hallucination with Dis tasks using Acc/P/R/F1/Specificity metrics.

Behind the Magic, MERLIM: Multi-modal Evaluation Benchmark for Large Image-Language Models

Evaluates visual grounding with Dis tasks using Accuracy metrics.

Unified Hallucination Detection for Multimodal Large Language Models

Evaluates t2i with Gen tasks using Acc/P/R/F metrics.

Hal-Eval: A Universal and Fine-grained Hallucination Evaluation Framework for Large Vision Language Models

Evaluates event hallucination with Dis & Gen tasks using Acc/Score metrics.

VALOR-EVAL: Holistic Coverage and Faithfulness Evaluation of Large Vision-Language Models

Evaluates multimodal hallucination with Gen tasks using Faithfulness & Coverage metrics.

AutoHallusion: Automatic Generation of Hallucination Benchmarks for Vision-Language Models

Evaluates multimodal hallucination with Dis tasks using ASR/MASR/CASR metrics.

ODE: Open-Set Evaluation of Hallucinations in Multimodal Large Language Models

Evaluates open-set with Dis & Gen tasks using AMBER/Acc metrics.

Measuring the Measurers: Quality Evaluation of Hallucination Benchmarks for Large Vision-Language Models

Evaluates ocr/action/counting with Dis & Gen tasks using Hallucination Rate metrics.

FIHA: Autonomous Hallucination Evaluation in Vision-Language Models with Davidson Scene Graphs

Evaluates eval method with Dis & Gen tasks using Acc/P/R/F1 metrics.

TUBench: Benchmarking Large Vision-Language Models on Trustworthiness with Unanswerable Questions

Evaluates unanswerable qs with Dis tasks using Acc metrics.

MFC-Bench: Benchmarking Multimodal Fact-Checking with Large Vision-Language Models

Evaluates manipulation/ooc/veracity with Dis & Gen tasks using Acc/F1 metrics.

Hallu-PI: Evaluating Hallucination in Multi-modal Large Language Models within Perturbed Inputs

Evaluates perturbed inputs with Dis tasks using Acc metrics.

CAST: Cross-modal Alignment Similarity Test for Vision Language Models

Evaluates self-consistency with Dis & Gen tasks using Similarity metrics.

FaithScore: Fine-grained Evaluations of Hallucinations in Large Vision-Language Models

Evaluates obj. counting with Gen tasks using FaithScore metrics.

ALOHa: A New Measure for Hallucination in Captioning Models

Evaluates captioning metric with Gen tasks using ALOHa metrics.

Does Object Grounding Really Reduce Hallucination of Large Vision-Language Models?

Evaluates metric method with Gen tasks using CHAIR-MEN / FaithScore metrics.

Visual Hallucinations of Multi-modal Large Language Models

Evaluates visual hallucination with Dis & Gen tasks using Acc metrics.

PhD: A ChatGPT-Prompted Visual hallucination Evaluation Dataset

Evaluates sentiment with Dis tasks using PhD Index metrics.

Automated Multi-level Preference for MLLMs

Evaluates multi-round conv with Gen tasks using Preference Score metrics.

Evaluating and Analyzing Relationship Hallucinations in Large Vision-Language Models

Evaluates multimodal hallucination with Dis tasks using Acc/P/R/F1 metrics.

Multi-Object Hallucination in Vision-Language Models

Evaluates multi-object with Gen tasks using Accuracy metrics.

BEAF: Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models

Evaluates before-after changes with Dis tasks using TU/IG/SB/ID metrics.

Antidote: A Unified Framework for Mitigating LVLM Hallucinations in Counterfactual Presupposition and Object Perception

Evaluates cp-bench mitigation with Gen tasks using Acc metrics.

EH-Benchmark: Ophthalmic hallucination benchmark and agent-driven top-down traceable reasoning workflow

Evaluates ophthalmology with Dis & Gen tasks using N/A metrics.

Evaluation and Analysis of Hallucination in Large Vision-Language Models

Evaluates multimodal hallucination with Gen tasks using LLM Rating metrics.

Holistic Analysis of Hallucination in GPT-4V(ision): Bias and Interference Challenges

Evaluates model bias with Gen tasks using Human Assessment metrics.

MMInstruct: A High-Quality Multi-Modal Instruction Tuning Dataset with Extensive Diversity

Evaluates instruction tuning with Gen tasks using Acc metrics.

MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation

Evaluates decoding method with Gen tasks using Acc metrics.

Unsolvable Problem Detection: Robust Understanding Evaluation for Large Multimodal Models

Evaluates unsolvable (upd) with Dis tasks using Acc metrics.

LOTUS: A Leaderboard for Detailed Image Captioning from Quality to Societal Bias and User Preferences

Evaluates societal bias/pref with Dis tasks using Alignment/Bias metrics.

How Easy is It to Fool Your Multimodal LLMs? An Empirical Analysis on Deceptive Prompts

Evaluates deceptive prompts with Gen tasks using Acc metrics.

VLind-Bench: Measuring Language Priors in Large Vision-Language Models

Evaluates language priors with Dis tasks using Acc (Y/N) metrics.

MMRel: Benchmarking Relation Understanding in Multi-Modal Large Language Models

Evaluates relation understanding with Dis & Gen tasks using Acc metrics.

Understanding Multimodal Hallucination with Parameter-Free Representation Alignment

Evaluates alignment method with Dis tasks using Pfram metrics.

GLSim: Detecting Object Hallucinations in LVLMs via Global-Local Similarity

A training-free detector that combines global scene similarity with local visual grounding for object hallucination detection.

Instruction Lens Score: Your Instruction Contributes a Powerful Object Hallucination Detector for Multimodal Large Language Models

A training-free object hallucination detector that uses calibrated local and instruction-context consistency scores.

Uncertainty Estimation in Autoregressive Structured Prediction

A general ensemble-based framework for token-level and sequence-level uncertainty estimation in autoregressive structured prediction.

Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT

ACT-ViT treats full layer-by-token activation tensors as image-like inputs for cross-LLM hallucination detection.

TruthPrInt: Mitigating Large Vision-Language Models Object Hallucination Via Latent Truthful-Guided Pre-Intervention

A latent-state detector identifies hallucinated object tokens and guides decoding toward a shared truthful direction.

Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations

A patch-level detector combines attention dispersion and cross-modal grounding consistency to localize hallucinated tokens.